Learning to Fuse Multiple Brain Functional Networks for Automated Autism Identification

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Functional connectivity networks (FCNs) are crucial for identifying brain dysfunction biomarkers like autism spectrum disorder (ASD).
• Existing FCN estimation methods capture only single relationships, failing to model complex brain interactions and often lacking task-specific optimization for ASD identification.

Purpose of the Study:

• To propose a multi-FCN fusion framework for enhanced rs-fMRI-based ASD classification.
• To address limitations of traditional unsupervised FCN estimation methods by incorporating label information for improved diagnostic performance.

Main Methods:

• Estimate multiple FCNs per subject using diverse methods to capture varied ROI interactions.
• Employ supervised learning with ASD/healthy control (HC) labels to derive fusion weights for optimal FCN importance.
• Apply the adaptively weighted fused FCN to the ABIDE dataset for ASD identification.

Main Results:

• The proposed multi-FCN fusion framework significantly improves diagnostic accuracy for ASD identification.
• The method outperforms traditional and state-of-the-art FCN estimation techniques.
• Achieved superior classification performance on the ABIDE dataset.

Conclusions:

• The multi-FCN fusion framework offers a straightforward yet effective approach for rs-fMRI-based ASD classification.
• Integrating multiple FCNs with adaptive weighting enhances the ability to identify complex brain alterations in ASD.
• This framework holds promise for improving diagnostic tools for neurodevelopmental disorders.